This patch adds two new kernels: SORT_BUCKET_PASS and SORT_WRITE_PASS. These replace PREFIX_SUM and SORTED_PATHS_ARRAY on supported devices (currently implemented on Metal, but will be trivial to enable on the other backends). The new kernels exploit sort partitioning (see D15331) by sorting each partition separately using local atomics. This can give an overall render speedup of 2-3% depending on architecture. As before, we fall back to the original non-partitioned sorting when the shader count is "too high".
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16909
This functionality is related only to debugging of SYCL implementation
via single-threaded CPU execution and is disabled by default.
Host device has been deprecated in SYCL 2020 spec and we removed it
in 305b92e05f.
Since this is still very useful for debugging, we're restoring a
similar functionality here through SYCL 2020 Host Task.
In this case the blocksize may not the one we requested, which was assumed to be
the case. Instead get the effective block size from the compiler as was already
done for Metal and OneAPI.
This patch adds a new Cycles device with similar functionality to the
existing GPU devices. Kernel compilation and runtime interaction happen
via oneAPI DPC++ compiler and SYCL API.
This implementation is primarly focusing on Intel® Arc™ GPUs and other
future Intel GPUs. The first supported drivers are 101.1660 on Windows
and 22.10.22597 on Linux.
The necessary tools for compilation are:
- A SYCL compiler such as oneAPI DPC++ compiler or
https://github.com/intel/llvm
- Intel® oneAPI Level Zero which is used for low level device queries:
https://github.com/oneapi-src/level-zero
- To optionally generate prebuilt graphics binaries: Intel® Graphics
Compiler All are included in Linux precompiled libraries on svn:
https://svn.blender.org/svnroot/bf-blender/trunk/lib The same goes for
Windows precompiled binaries but for the graphics compiler, available
as "Intel® Graphics Offline Compiler for OpenCL™ Code" from
https://www.intel.com/content/www/us/en/developer/articles/tool/oneapi-standalone-components.html,
for which path can be set as OCLOC_INSTALL_DIR.
Being based on the open SYCL standard, this implementation could also be
extended to run on other compatible non-Intel hardware in the future.
Reviewed By: sergey, brecht
Differential Revision: https://developer.blender.org/D15254
Co-authored-by: Nikita Sirgienko <nikita.sirgienko@intel.com>
Co-authored-by: Stefan Werner <stefan.werner@intel.com>
* Rename "texture" to "data array". This has not used textures for a long time,
there are just global memory arrays now. (On old CUDA GPUs there was a cache
for textures but not global memory, so we used to put all data in textures.)
* For CUDA and HIP, put globals in KernelParams struct like other devices.
* Drop __ prefix for data array names, no possibility for naming conflict now that
these are in a struct.
Move MNEE to own kernel, separate from shader ray-tracing. This does introduce
the limitation that a shader can't use both MNEE and AO/bevel, but that seems
like the better trade-off for now.
We can experiment with bigger kernel organization changes later.
Differential Revision: https://developer.blender.org/D15070
This patch makes it possible to change the precision with which to
store volume data in the NanoVDB data structure (as float, half, or
using variable bit quantization) via the previously unused precision
field in the volume data block.
It makes it possible to further reduce memory usage during
rendering, at a slight cost to the visual detail of a volume.
Differential Revision: https://developer.blender.org/D10023
Found those missing casts while looking into a crash report made in
the Blender Chat. Was unable to reproduce the crash, but the casts
should totally be there to avoid integer overflow.
To make porting to other architectures easier, clarifying that this does not
need to be supported. The unused parallel_reduce implementation assumed warp
size 32, but is easy to update if we ever need it in the future.
Workaround for a compilation issue preventing kernels compiling for AMD GPUs: Avoid problematic use of templates on Metal by making `gpu_parallel_active_index_array` a wrapper macro, and moving `blocksize` to be a macro parameter.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D14081
* Replace license text in headers with SPDX identifiers.
* Remove specific license info from outdated readme.txt, instead leave details
to the source files.
* Add list of SPDX license identifiers used, and corresponding license texts.
* Update copyright dates while we're at it.
Ref D14069, T95597
This patch fixes a couple of new Metal kernel compilation errors: 1) a kernel parameter count overflow, and 2) missing address space qualifiers.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D13763
Enables the `bpy.ops.cycles.denoise_animation()` operator again and modifies it to support
temporal denoising with OptiX. This requires renders that were done with both the "Vector"
and "Denoising Data" passes.
Differential Revision: https://developer.blender.org/D11442
This patch adds the Metal host-side code:
- Add all core host-side Metal backend files (device_impl, queue, etc)
- Add MetalRT BVH setup files
- Integrate with Cycles device enumeration code
- Revive `path_source_replace_includes` in util/path (required for MSL compilation)
This patch also includes a couple of small kernel-side fixes:
- Add an implementation of `lgammaf` for Metal [Nemes, Gergő (2010), "New asymptotic expansion for the Gamma function", Archiv der Mathematik](https://users.renyi.hu/~gergonemes/)
- include "work_stealing.h" inside the Metal context class because it accesses state now
Ref T92212
Reviewed By: brecht
Maniphest Tasks: T92212
Differential Revision: https://developer.blender.org/D13423
This patch adds MetalRT support to Cycles kernel code. It is mostly additive in nature or confined to Metal-specific code, however there are a few areas where this interacts with other code:
- MetalRT closely follows the Optix implementation, and in some cases (notably handling of transforms) it makes sense to extend Optix special-casing to MetalRT. For these generalisations we now have `__KERNEL_GPU_RAYTRACING__` instead of `__KERNEL_OPTIX__`.
- MetalRT doesn't support primitive offsetting (as with `primitiveIndexOffset` in Optix), so we define and populate a new kernel texture, `__object_prim_offset`, containing per-object primitive / curve-segment offsets. This is referenced and applied in MetalRT intersection handlers.
- Two new BVH layout enum values have been added: `BVH_LAYOUT_METAL` and `BVH_LAYOUT_MULTI_METAL_EMBREE` for XPU mode). Some host-side enum case handling has been updated where it is trivial to do so.
Ref T92212
Reviewed By: brecht
Maniphest Tasks: T92212
Differential Revision: https://developer.blender.org/D13353
This patch fixes an address space mismatch in the film convert kernels on Metal. The `film_get_pass_pixel_...` functions take a `ccl_private` result pointer, but the film convert kernels pass a `ccl_global` memory pointer. Specialising the pass-fetch functions with templates results in compilation errors on Visual Studio, so instead this patch just adds an intermediate local on Metal.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D13350
With the current code in master, scrambling distance is enabled on non-hardware accelerated ray tracing devices see a measurable performance decrease when compared scrambling distance on vs off. From testing, this performance decrease comes from the large tile sizes scheduled in `tile.cpp`.
This patch attempts to address the performance decrease by using different algorithms to calculate the tile size for devices with hardware accelerated ray traversal and devices without. Large tile sizes for hardware accelerated devices and small tile sizes for others.
Most of this code is based on proposals from @brecht and @leesonw
Reviewed By: brecht, leesonw
Differential Revision: https://developer.blender.org/D13042
MSL requires that constant address space literals be declared at program
scope. This patch moves the `blackbody_table_r/g/b` and `cie_colour_match`
constants into separate files so they can be declared at the appropriate scope.
Ref T92212
Differential Revision: https://developer.blender.org/D13241
This patch contains many small leftover fixes and additions that are
required for Metal-enablement:
- Address space fixes and a few other small compile fixes
- Addition of missing functionality to the Metal adapter headers
- Addition of various scattered `__KERNEL_METAL__` blocks (e.g. for
atomic support & maths functions)
Ref T92212
Differential Revision: https://developer.blender.org/D13263
The issue was caused by splitting happening twice.
Fixed by checking for split flag which is assigned to the both states
during split.
The tricky part was to write catcher data at the moment of split: the
transparency and shadow catcher sample count is to be accumulated at
that point. Now it is happening in the `intersect_closest` kernel.
The downside is that render buffer is to be passed to the kernel, but
the benefit is that extra split bounce check is not needed now.
Had to move the passes write to shadow catcher header, since include
of `film/passes.h` causes all the fun of requirement to have BSDF
data structures available.
Differential Revision: https://developer.blender.org/D13177
It's unclear why this fails. Maybe the size of half4 is not the expected
8 bytes and adjacent pixels are overwritten. Or there is some bug in the
HIP compiler writing a struct into global memory, which we probably don't
do elsewhere in the kernel.
Thanks to Thomas, William and Jeroen for helping investigate this.
rB3a4c8f406a3a3bf0627477c6183a594fa707a6e2 changed the macros that create the film
convert kernel entry points, but in the process accidentally changed the parameter definition
to one of those (which caused CUDA launch and misaligned address errors) and changed the
implementation as well. This restores the correct implementation from before.
In addition, the `ccl_gpu_kernel_threads` macro did not work as intended and caused the
generated launch bounds to end up with an incorrect input for the second parameter (it was
set to "thread_num_registers", rather than the result of the block number calculation). I'm
not entirely sure why, as the macro definition looked sound to me. Decided to simply go with
two separate macros instead, to simplify and solve this.
Also changed how state is captured with the `ccl_gpu_kernel_lambda` macro slightly, to avoid
a compiler warning (expression has no effect) that otherwise occurred.
Maniphest Tasks: T92985
Differential Revision: https://developer.blender.org/D13175
This patch adapts the shared kernel entrypoints so that they can be compiled as MSL (Metal Shading Language). Where possible, the adaptations avoid changes in common code.
In MSL, kernel function inputs are explicitly bound to resources. In the case of argument buffers, we declare a struct containing the kernel arguments, accessible via device pointer. This differs from CUDA and HIP where kernel function arguments are declared as traditional C-style function parameters. This patch adapts the entrypoints declared in kernel.h so that they can be translated via a new `ccl_gpu_kernel_signature` macro into the required parameter struct + kernel entrypoint pairing for MSL.
MSL buffer attribution must be applied to function parameters or non-static class data members. To allow universal access to the integrator state, kernel data, and texture fetch adapters, we wrap all of the shared kernel code in a `MetalKernelContext` class. This is achieved by bracketing the appropriate kernel headers with "context_begin.h" and "context_end.h" on Metal. When calling deeper into the kernel code, we must reference the context class (e.g. `context.integrator_init_from_camera`). This extra prefixing is performed by a set of defines in "context_end.h". These will require explicit maintenance if entrypoints change. We invite discussion on more maintainable ways to enforce correctness.
Lambda expressions are not supported on MSL, so a new `ccl_gpu_kernel_lambda` macro generates an inline function object and optionally capturing any required state. This yields the same behaviour. This approach is applied to all parallel_... implementations which are templated by operation. The lambda expressions in the film_convert... kernels don't adapt cleanly to use function objects. However, these entrypoints can be macro-generated more concisely to avoid lambda expressions entirely, instead relying on constant folding to handle the pixel/channel conversions.
A separate implementation of `gpu_parallel_active_index_array` is provided for Metal to workaround some subtle differences in SIMD width, and also to encapsulate some required thread parameters which must be declared as explicit entrypoint function parameters.
Ref T92212
Reviewed By: brecht
Maniphest Tasks: T92212
Differential Revision: https://developer.blender.org/D13109
Adds a pass before denoising that calculates the intensity of the image, which can be
passed into the OptiX denoiser for more optimal results for very dark or very bright images.
In addition this also fixes a crash that sometimes occurred on exit. The OptiX denoiser object
has to be destroyed before the OptiX device context object (since it references that). But in
C++ the destructor function of a class is called before its fields are destructed, so
"~OptiXDevice" was always called before "OptiXDevice::~Denoiser" and therefore
"optixDeviceContextDestroy" was called before "optixDenoiserDestroy", hence the crash.
Differential Revision: https://developer.blender.org/D13160
Remove prefix of filenames that is the same as the folder name. This used
to help when #includes were using individual files, but now they are always
relative to the cycles root directory and so the prefixes are redundant.
For patches and branches, git merge and rebase should be able to detect the
renames and move over code to the right file.
* Split render/ into scene/ and session/. The scene/ folder now contains the
scene and its nodes. The session/ folder contains the render session and
associated data structures like drivers and render buffers.
* Move top level kernel headers into new folders kernel/camera/, kernel/film/,
kernel/light/, kernel/sample/, kernel/util/
* Move integrator related kernel headers into kernel/integrator/
* Move OSL shaders from kernel/shaders/ to kernel/osl/shaders/
For patches and branches, git merge and rebase should be able to detect the
renames and move over code to the right file.
Similar to main path compaction that happens before adding work tiles, this
compacts shadow paths before launching kernels that may add shadow paths.
Only do it when more than 50% of space is wasted.
It's not a clear win in all scenes, some are up to 1.5% slower. Likely caused
by different order of scheduling kernels having an unpredictable performance
impact. Still feels like compaction is just the right thing to avoid cases
where a few shadow paths can hold up a lot of main paths.
Differential Revision: https://developer.blender.org/D12944
Taking advantage of the new decoupled main and shadow paths. For CPU we
just store two nested structs in the integrator state, one for direct light
shadows and one for AO. For the GPU we restrict the number of shade surface
states to be executed based on available space in the shadow paths queue.
This also helps improve performance in benchmark scenes with an AO pass,
since it is no longer needed to use the shader raytracing kernel there,
which has worse performance.
Differential Revision: https://developer.blender.org/D12900
These transparent shadows can be expansive to evaluate. Especially on the
GPU they can lead to poor occupancy when only some pixels require many kernel
launches to trace and evaluate many layers of transparency.
Baked transparency allows tracing a single ray in many cases by accumulating
the throughput directly in the intersection program without recording hits
or evaluating shaders. Transparency is baked at curve vertices and
interpolated, for most shaders this will look practically the same as actual
shader evaluation.
Fixes T91428, performance regression with spring demo file due to transparent
hair, and makes it render significantly faster than Blender 2.93.
Differential Revision: https://developer.blender.org/D12880
The motivation for this is twofold. It improves performance (5-10% on most
benchmark scenes), and will help to bring back transparency support for the
ambient occlusion pass.
* Duplicate some members from the main path state in the shadow path state.
* Add shadow paths incrementally to the array similar to what we do for
the shadow catchers.
* For the scheduling, allow running shade surface and shade volume kernels
as long as there is enough space in the shadow paths array. If not, execute
shadow kernels until it is empty.
* Add IntegratorShadowState and ConstIntegratorShadowState typedefs that
can be different between CPU and GPU. For GPU both main and shadow paths
juse have an integer for SoA access. Bt with CPU it's a different pointer
type so we get type safety checks in code shared between CPU and GPU.
* For CPU, add a separate IntegratorShadowStateCPU struct embedded in
IntegratorShadowState.
* Update various functions to take the shadow state, and make SVM take either
type of state using templates.
Differential Revision: https://developer.blender.org/D12889
* Rename struct KernelGlobals to struct KernelGlobalsCPU
* Add KernelGlobals, IntegratorState and ConstIntegratorState typedefs
that every device can define in its own way.
* Remove INTEGRATOR_STATE_ARGS and INTEGRATOR_STATE_PASS macros and
replace with these new typedefs.
* Add explicit state argument to INTEGRATOR_STATE and similar macros
In preparation for decoupling main and shadow paths.
Differential Revision: https://developer.blender.org/D12888
